TITLE:
PsychoGraph-Net: A Graph-Attentive Deep Learning Framework for Mapping, Predicting, and Interpreting Psychiatric Comorbidity Networks in Liaison Psychiatry
AUTHORS:
Rocco de Filippis, Abdullah Al Foysal
KEYWORDS:
Psychiatric Comorbidity, Graph Attention Network, Liaison Psychiatry, Network Science, Bayesian Deep Learning, Electronic Health Records, Community Detection, Computational Psychiatry, ICD-10, Machine Learning, Integrated Care
JOURNAL NAME:
Open Access Library Journal,
Vol.13 No.5,
May
28,
2026
ABSTRACT: Psychiatric comorbidity in liaison psychiatry is not random noise, it follows structured, clinically meaningful patterns that standard episodic assessment cannot detect. Identifying these patterns computationally could transform how clinicians’ reason about complex inpatients who carry multiple co-occurring diagnoses. We introduce PsychoGraph-Net, an end-to-end graph-attentive deep learning framework that ingests electronic health record (EHR) data, constructs a weighted psychiatric comorbidity graph from ICD-10-coded diagnoses, and jointly performs 1) community detection to discover diagnostically coherent clusters and 2) five-class mood-state classification (Euthymia, Depression, Hypomania, Mania, Mixed Features) as the primary prediction endpoint, with episode-onset risk at the 7-day horizon as a secondary endpoint, both implemented via a Graph Attention Network with Bayesian uncertainty quantification. Built on a retrospective cohort of 14,287 hospitalized patients from a tertiary liaison psychiatry service over six years, the resulting comorbidity graph comprises 89 diagnostic nodes and 1247 clinically significant co-occurrence edges. PsychoGraph-Net achieves an AUC-ROC of 0.951, macro-F1 of 0.887, and an Expected Calibration Error (ECE) of 0.031 outperforming all evaluated baselines by statistically significant margins. Community detection reveals three previously uncharacterized comorbidity clusters: (A) an Affective-Anxiety-Sleep cluster centred on major depressive disorder, (B) a Neurocognitive-Delirium complex anchored by delirium and dementia, and (C) a Somatic-Metabolic interface linking somatic symptom disorders with cardiovascular and metabolic disease. Bayesian uncertainty flagging routes 14.7% of high-uncertainty predictions to clinician review, reducing diagnostic workload by 31.4% in clinical simulation without sensitivity loss. PsychoGraph-Net is the first validated, end-to-end graph neural network designed specifically for psychiatric comorbidity analysis in liaison psychiatry.